3.1 Network architecture
Because UPIoT tends to cover a relatively wide area and produce massive and distributed data, signal processing and data analytics theories and techniques are needed to handle the data and observe the state of the large-scale system on which it resides. The distributed signal processing theories of UPIoT, which include distributed sensing and cooperative estimation/detection, are discussed here.
The signal processing architecture mainly includes centralized architecture, decentralized architecture, distributed architecture, and hybrid architecture [12, 13]. The optimal architecture is selected according to different objectives as follows:
-
1)
Centralized architecture: It is the most common architecture in practical application, which is relatively simple in structure and relatively mature in algorithms and techniques but has high requirements for bandwidth and computing power.
-
2)
Decentralized architecture: It is used when the system consists of different sub-systems. The sensor nodes in the sub-systems communicate with each other, and the fusion processing centers in each sub-system also communicate with others in different sub-systems. Thus, the communication burden across the sub-systems is reduced, but the internal communication burden within each sub-system is heavy.
-
3)
Distributed architecture: There is no local fusion processing center, and each sensor node transmits the estimation results to the adjacent sensor node, which are also used to correct the estimation results. The advantages are that the communication cost is low, and the network scalability is good, and the edge computing may play an important role in this situation.
-
4)
Hybrid architecture: The hybrid computing architecture with overall decentralized and locally distributed processing can also be employed.
3.2 Theories of distributed sensing and cooperative estimation/detection
3.2.1 Distributed sensing
In distributed sensing [14], a group of sensors collectively observe information of the state of the environment. Due to factors such as cost, spectrum bandwidth limitation, and complexity, the sensors compress their own observation data and transmit only part of the information to the fusion center to estimate the state of the environment. Compared with centralized sensing, distributed sensing will lose some measurement information, and consequently, unprocessed sensing information can make the fusion center difficult to handle in large-scale systems, leading to real-time performance deterioration or even collapsing.
Distributed sensor systems have many different types of network structures. They usually form a connected graph with no loops, so information transmits along the uniquely determined path from the sensor node to the fusion center. Another type is the parallel structure in which each sensor node communicates directly with the fusion center. There is also a random configuration with sensor nodes randomly placed in the area cooperating with each other to form a dynamic communication network. The most common data transmission models of wireless sensor networks (WSNs) are shown in Fig. 5 [15].
The ideal distributed perception architecture is that each node in the network has equal status and can be used as a sensor, a router, a repeater, and so on [12]. A sensor located in the target event area senses its information and transmits the information to the fusion center through a single hop or multiple hops. Other nodes can be used as repeaters or routers, and the fusion center estimates the final results after one or more data transmissions.
3.2.2 Cooperative estimation/detection
The commonly used estimation methods for static estimation (Fig. 6) are the moment estimation method, maximum likelihood estimation, least squares method, Bayesian estimation, and so on. The most commonly used estimation method for dynamic estimation is the Kalman filtering method [16].
The distributed feature of UPIoT is that different sensors jointly collect state information in an area. Because of bandwidth and power limitations, the observed information needs to be quantized, and the collected signal is compressed. According to existing studies, the challenges of UPIoT-based estimation can be summarized as follows [12, 1) how to reduce the dimension to compress the observation data, while maintaining the mean square error; (2) how to quantize the compressed signal before transmission; and (3) how to construct an estimator to quantize the compressed information. The theories for data compression and cooperative detection/estimation used in UPIoT are further introduced here.
-
A.
Data compression
With the development of Internet Plus Smart Energy, the scale of the system and the requirements for communication systems are getting increasingly larger and higher. Thus, UPIoT may face the problem of insufficient computing power, storage capacity, transmission capacity, and power supply capacity. How to compress information and reduce the communication burden has become one of the key issues in current research [17, 18].
At present, the sparse technology commonly used in complex power systems reduces storage pressure by only storing non-zero elements, but as the system scale increases, sparse technology cannot meet the demand. In 2006, Candes et al. proposed a compressed sensing technology with which the signal is sparse or can be sparsely represented after transformation and can be sampled with a sampling frequency far lower than the Nyquist signal to obtain a compressed signal [19]. Therefore, perception and compression are completed together in one step. The comparison of the two processes is shown in Fig. 7.
-
B.
Cooperative detection and cooperative estimation
In an N-node system, assuming a change at time λ = k, the observed values \( {X}_1^i,\cdots, {X}_{k-1}^i \) on sensor i are independent and identically distributed, whereas observation \( {X}_k^i,{X}_{k+1}^i \) follows another distribution. The observations of each sensor are independent of each other. When t = n, the observation information for sensor node i is \( {X}_n^i=\left[{X}_1^i,\cdots, {X}_n^i\right] \), and the detection problem can be equivalent to the hypothesis test problem of “H0: λ > n” and “H1: λ ≤ n”. If H0 is true, no change occurs, and if H0 is false, change occurs. In cooperative detection, in the absence of a fusion or control center, each sensor detects changes within the region in a parallel or distributed manner and interacts with information through a certain data transfer model. Finally, based on the information collected by itself and the interactive information, the local system can be detected. When there is a fusion or control center, the observations of all sensors are available at the fusion or control center. The cooperative detection framework is shown in Fig. 8.
After detecting changes in the system, it is necessary to estimate the observations to obtain the system status. Cooperative estimation means that in distributed networks the nodes exchange information with each other and update and correct their own information according to the exchanged information [20]. Cooperative estimation has the following advantages over centralized estimation [21,22,23]: cooperative estimation of each node’s independent information and exchanging information with neighboring nodes can effectively reduce the computational time and increase the robustness of the network.
Considering a WSN system consisting of N sensors, here we try to estimate a deterministic parameter S. Since the fusion center receives the quantitative observations, the goal is to obtain \( \hat{S} \) with the least square error based on this quantitative observation. The nth sensor can observe that
$$ x(n)=s+w(n),n\in \left(0,N-1\right) $$
(1)
where w(n) is Gaussian white noise, which obeys a (0, σ2) distribution, and the w(n) functions of different sensors are independent and identically distributed.
The quantified indicators are constructed as
$$ {b}_k(n)=1\left\{x(n)\in {B}_k(n)\right\},k\in \left[1,K\right] $$
(2)
which takes a value of one when x(n) ∈ Bk(n) ⊂ RM is satisfied; otherwise, it takes a value of zero.
After the signal is quantized, the network transmission path is used to transmit the quantized information to other nodes. After the information is exchanged, the node information is updated, and a suitable distributed algorithm is selected to cooperatively estimate the deterministic parameter S.
The ideal distributed estimation has the following characteristics [24]:
-
1)
Accurate estimation. After partitioning the geographically distributed large-scale system, even when the system measurement value is not objective, each area can still accurately estimate the system operation status.
-
2)
Small communication burden so real-time requirements can be met.
-
3)
Distributed architecture with accuracy like that of the centralized estimation.
-
4)
Good real-time performance due to the small amount of calculation required.
-
5)
High robustness.